Yeah, they got that 'no pores' look that selfie filter give, that's somewhat uncanny looking.

Human male 40 absolutely does not have a filter, that's probably why he rated as the most human looking human.

Not only that but several of them are a bit weird looking (sorry to those people...) as in, 37 and 47 have obvious asymmetries, 31 is a bit bug-eyed, 18 seems to have been taken with a super telephoto lens or have a really flat face.

This reminds me of an argument I saw here last week about AI and its use as a grammar checker. You can definitely do it, but you’re going to have all the markers of using AI to cheat.

An op? Making a misleading title? On Lemmy?

Man, it's as if the severe lack of moderation and rules that so many people wanted when moving from Reddit is hurting the quality of posts on here.

@cyu Could you add “white” between “actual” and “human”?

It turns out that AI faces were rated as more human-like than actual humans

I tried guessing from the ArsTechnica article and got a whopping 1 out of 8 correct.

It’s the “living on Null street” problem of the future

His skin is simply too smooth, plus his eyes are at different angles and the perspective is wrong

Not really much of an issue with SXDL any more, and even SD1.5 got quite good at the end. Admittedly haven't stress-tested either, though (things like clasping hands etc). It's also not a hand-specific thing, it just happens more commonly with fingers because they're small features:

The thing that happens is that diffusion-type interference first nails down gross structure (which limb is where) and then fills in details. Sometimes steps somewhere in the middle decide that a limb should be somewhere else, though, and suddenly you have two, and if steps immediately after don't think "that old limb doesn't look like it should be there" and erase it, later steps will happily refine both to photorealism because they don't even look at the overall composition. That is, it's not an issue with anatomical knowledge, or not having seen enough hands, but the model changing its mind but not backtracking. It's actually astonishing how good it can get at not making that mistake without being able to tell that it has two competing goals in mind.

Such is the miracle of sufficiently trained autoencoders

It's a valley. Meaning you be either before, in, or after. We are after

We used the 100 AI and 100 human White faces (half male, half female) from Nightingale and Farid. The AI faces were generated using StyleGAN2. The human faces were selected from the Flickr-Faces-HQ Dataset to match each of the AI faces as closely as possible (e.g., same gender, posture, and expression). All stimuli had blurred or mostly plain backgrounds, and AI faces were screened to ensure they had no obvious rendering artifacts (e.g., no extra faces in background). Screening for artifacts mimics how real-world users screen AI faces, either as scientists or for public use, and therefore captures the type and range of stimuli that appear online. Participants were asked to resize their screen so that stimuli had a visual angle of 12° wide × 12° high at ~50 cm viewing distance.

She looks great for 44.

This is a classification made by humans, not by AI

Exactly. The AIs job is to generate humanness. The things that don’t look human get discarded, the things that have strong human indicators get kept. Oh look, the AI did its job. Shocked pikachu.

The white thing is probably just a case of biased training data. Which is going to be a problem across all AIs. I wouldn’t be surprised if in 5-10 years (if the fad lasts longer than NFTs lmao) we find out the ‘AIs’ have all been fed biased data as yet another means of large corporations controlling the narrative of the population.